With the recent, rapid development of computer tools, digital image processing has become an efficient, economic, and convenient means for obtaining qualitative and quantitative information in different fields such as remote sensing, cartography, robotics, and materials.
Indeed, digital image processing makes it possible to qualitatively describe images from various sources in terms of pattern recognition to identify and isolate contained objects. One of the major subclasses of contained objects present in digital images consists in linear features. Automatic detection of linear features from digital images plays an important role in pattern recognition and in digital image processing.
A variety of techniques and many algorithms have emerged to automatically extract linear features from digital images. These techniques can be classified into two main categories: local methods which are based on local operators such as mobile kernel, and global methods which focus on mathematical transformations, such as Hough transforms (see [1]).
Local methods for automatically extracting linear features basically exploit local variations of pixel intensity in a small neighborhood by calculating the gradients in small, limited-size windows in the image, e.g., 3×3 pixels or 5×5 pixels (see [2][3][4][5][6][7][8][9]).
A number of researchers have examined mathematical morphology as a means of extracting linear features (see [10][11][12][13][14]). The problems pertaining to this technique arise from the number of human decisions required to reconnect and rebuild line segments, which increases processing time. Multi-dimensional line detection is the other technique for detecting for linear features that collects different spectral information for the same scene and may highlight different parts of lines (see [15]). The first stage of this method for obtaining the combined images requires several transformations of multiple original bands. Human intervention is needed to select the best-combined image. Another approach to linear-feature extraction involves knowledge-based systems, which need more information than a simple digital image for line extraction (see [16][17][18]).
These local methods generally remain inefficient because they fail to have a global view of the linear features in a digital image. One problem common to all of these methods is that the resulting extracted line images contain a fair amount of noise, while the detected lines are often incomplete and geometrically shifted. These difficulties are magnified by intersections between different linear features and linear features that display some curvature (see [19][20][1]). In addition, these methods turn in exceptionally long processing time when extracting features from large images (see [16]).
Of the local methods, LINDA (Linear-featured Network Detection and Analysis) system (see [2][7][27]), based on the profile intensity analysis of the pixel line, is the most recent and the most representative method. With all of basic disadvantages of the local methods, the LINDA system is far from being operational with respect to systematically and automatically processing a large set of images.
The Radon transform and its derivative, the Hough transform, are the most frequently used approaches as global methods for detecting linear features (see [21][22][23][24]). In principle, a straight line from the input image is transformed into a digital peak (a light or dark pixel, compared to its neighborhood) in the transformed plane. In this case, it is easier to detect a peak in the transformed plane than a straight line in the input image. There are three basic limitations for these methods that unfortunately restrict their applications and their utility in practice.
First, the Hough and Radon transform-based methods are limited solely to extracting straight lines (see [21]). Therefore, linear features that span the entire image but display some curvature may not produce suitable peaks or troughs in the transform plane. This restriction is linked directly to the basic definition of the method.
Secondly, there are false peaks or virtual maxima (peaks that do not represent the real line on the input image) in the transformed plane. These false peaks considerably decrease the quality of the results by increasing the error of commission. When the line density in the input image is high, eliminating the false peaks from transformed plane seriously limits these methods.
The last methodological inconvenience of these approaches is the ignorance of the nature of the detected lines. Indeed, since the intensity integration in the transformation process is performed over the entire length of the image, the process can have difficulty detecting line segments that are significantly shorter than the image dimensions (see [21][25][26]). Neither can it provide information about the positions of the endpoints of these shorter line segments or line length.
It is to be noted that the present description refers to a number of documents indicated in brackets, that are listed hereinbelow, the content of these documents is herein incorporated by reference in their entirety.